Research Of Cross Domain Sentiment Classification For Product Reviews

With the development of information technology, more and more user sentiment data are available from websites such as weibo, product reviews, film comments and user feedback. These comment and feedback data involve many domains including books, computers and cosmetics. It will help the enterprise master the user attitude of products in time by providing the automatic prediction of sentiment polarity. Furthermore, it will help to analyze the user's demand and provide the basis for personalized recommendation. However, with the increasing of applications in product fields of automatic sentiment analysis, the shortage of high cost of manual labeling and the lack of samples in some practical applications becomes higher in the traditional single domain learning based approaches. Considering the similar characteristics in expressing the sentiment with text, researchers begin to explore the methods which use the sentiment analysis model trained in a certain domain to the other domains. However, different words are used in expressing the same attitude in different domains in the real life, so the classifier trained in certain domain can not be applied to other domains directly. Thus it becomes an important research topic to propose some effective cross-domain sentiment classification methods by employing the knowledge transfer from one domain to the others.The key of cross-domain sentiment analysis is to reduce the gap between the domains. The current approaches are usually to explore the adaptive solutions in three aspects, such as instance adaptation, feature adaptation and model adaptation. The feature adaptation approaches are based on the feature representation transfer, in which the data in the source domain and target domain are mapped into the same feature distribution through feature representation. Compared with the methods of instance adaptaion, feature adaptation can establish the mapping relationship between domain words in a more abstract level. With the help of shared words as a bridge, the spectral feature alignment can reduce the gap between domains and discover a robust representation for cross-domain data. Based on the above approach, this paper further explores the relationship between shared words and domain-special words, and proposes the word alignment method based on association rules. In addition, the paper discusses the adavantages of deep learning in extracting abstract feature, and improves the deep learning method in text sentiment analysis and further explores the domain adaptation methods in model level. The effectiveness of the proposed method is verified on the public datasets. The main research work and innovations are introduced as follows:(1) This paper proposes a word alignment method based on association rules to achieve the word mapping. The proposed method enhances the alignment effect of the different words in different domains which express the same emotion and provides the basis for cross-domain sentiment classification. This is achieved by using the shared words as a bridge to reduce the gap between domains, discovering the rules between shared words and domain-special words in the same domain, and aligning the domain-special words from different domains to construct same feature space. The experiment results on the benchmark dataset from customer reviews about four category products on Amazon show that our proposed method achieves competitive accuracy, transfer loss and transfer ratio with the state-of-the-art methods of cross-domain sentiment classification.(2) This paper proposes a dynamic convolutional extreme learning machine algorithm for text sentiment classification, where the extreme learning machine is used to improve the generalization and effectiveness of the former dynamic convolutional neural networks in text sentiment classification. The proposed algorithm has taken the advantages of convolution operation in extracting the salient features from text vectors. Moreover, to improve the universal property of the classification layer, we propose to modify the fully connection layer with the single-layer network of extreme learning machine. By utilizing the perturbation ability of the random generation of parameters, it is prone to mitigate the dependence on training samples and avoid over-fitting. Experiments on several public data sets show that our proposed approach outperforms the dynamic convolutional neural network and some extended algorithms of extreme learning machine under the evaluation metrics including accuracy rate, F1-measure etc.(3) Study the characteristics of dynamic convolutional extreme learning machine algorithm in cross-domain sentiment classification, and explore the improvement and application methods of cross-domain dynamic convolution extreme leaning model. The model uses the labeled samples in the source domain to learn the parameters of deep network model, and applies the learned network parameters before the fully connection layer to extract the feature representation of the target domain data. Then a small amount of labeled samples in the target domain are used to train the adaptation layer. Experiments for cross-domain sentiment classification on Amazon data sets show that our proposed model is better than some relative algorithms which have no knowledge transfer under the accuracy rate, and has the certain ability of domain adaptation.